基于自相似性上下文和混合注意力的无监督可变形医学图像配准-《中国医学物理学杂志》

文章信息/Info

Title:: Unsupervised deformable medical image registration based on self-similarity context and mixed attention

作者:: 李碧草 1; 王岩 1; 王贝 2; 邵珠宏 3; 郭旭伟 4; 衣本泽 1; 1. 中原工学院信息与通信工程学院,河南郑州 450007;2. 中原工学院校医院,河南郑州 451191;3. 首都师范大学信息工程学院,北京 100048;4. 河南科技大学第一附属医院儿科,河南洛阳 471000

Author(s):: LI Bicao1; WANG Yan1; WANG Bei2; SHAO Zhuhong3; GUO Xuwei4; YI Benze1; 1. School of Electronic and Information Engineering, Zhongyuan University of Technology, Zhengzhou 450007, China; 2. Infirmary, Zhongyuan University of Technology, Zhengzhou 451191, China; 3. Information Engineering College, Capital Normal University, Beijing 100048, China; 4. Department of Pediatrics, the First Affiliated Hospital of He’nan University of Science and Technology, Luoyang 471000, China

Keywords:: deformable medical image registration; self-similarity context; mixed attention; unsupervised deep learning

摘要:: 为了充分利用 Transformer 进行精确的配准,采用自相似性上下文作为特征提取器提取体素邻域上下文的语义信息。它使用具有扩散正则化的对称多尺度离散优化来寻找平滑的变换,可以快速地计算描述符之间的逐点距离。此外, 提出一种基于混合注意力的 Transformer 网络(STWA),结合通道、空间注意力以及基于(移动)窗口的自注意力方案,充分利用 3 种注意力机制的互补优势,既能利用全局统计信息,又具有强大的局部拟合能力。在 LPBA40、IXI 和 OASIS 3 个 3D 大脑 MRI 数据集上进行全面的实验,结果表明,与常用的配准方法 SyN、VoxelMorph、CycleMorph、ViT-V-Net 和 TransMorph 相比,本文方法在评估指标上实现优越的性能,证明模型在可变形医学图像配准中的有效性。

Abstract:: To fully exploit Transformer for accurate registration, self-similarity context is used as a feature extractor to extract the semantic information of the voxel neighborhood context, using symmetric multi-scale discrete optimization with diffusion regularization to find smooth transformations for quickly calculating the point-by-point distance between descriptors. In addition, a spatial-channel Transformer based on window attention network is proposed, which combines channel, spatial attention and self-attention scheme based on (moving) window, and makes full use of the complementary advantages of these 3 attention mechanisms, enabling the network to utilize global statistical information and have strong local fitting ability. The results of comprehensive experiments on 3D brain MRI datasets of LPBA40, IXI and OASIS shows that the proposed method is superior to the commonly used registration methods (SyN, VoxelMorph, CycleMorph, ViT-V-Net and TransMorph) on several evaluation indicators, proving its effectiveness in deformable medical image registration.